Professional Documents
Culture Documents
1. Introduction
This paper describes continued work in a new kind of decision
systemone based on emotions or affect and whose implementation is directly modeled on and constrained by those systems in the mammalian brain that are presently thought to handle affect. The underlying premise is that humans (and all
mammals) emotionally evaluate actions in terms of how these
will influence their well-being and that consequently, being
able to reproduce these affect-based evaluations in autonomous
agents will provide for an additional method of making and/or
supporting decisions. We intentionally constrain ourselves to
the underlying empirical evidence from mainstream neuroscience and adopt the philosophical stance that when presented
with two approaches, we limit ourselves to that which has a
sounder basis in the underlying biological empirical evidence.
The approach offers the promise of advancing understanding
and capability in three areas: autonomous decision-making,
semantics, and affect modeling. Using a search heuristic in
which the external environment modifies internal affect with
the aim of maximizing internal positive affect provides the basis of an autonomous decision engine, which, because of its explicitly non-cognitive bias, is fast enough for use in real-time
applications. The affect generated in the agent by interacting
with objects in its environment in terms of the impact of these
on the agents well-being, promises an affect-based approach to
semantics. The modeling of affect will also advance under-
Previous work [18, 19, 20], proposed that a major part of decision-making involves emotion-based evaluations of objects
and ideas. We proposed an architecture that follows closely
findings in neuroscience. Our goal is to articulate and validate
the concept that non-conscious software agents can develop a
sense of value of the objects with which they interact. This
evaluation is in terms of how these objects influence the agents
well-being, i.e., its ability to function effectively. A key measure of well-being is the agents affective state.
Such an approach promises the ability to create systems
that can interact with real-world objects with some understanding of what these objects mean that is over and above analytical
understandings typical of standard computing. The promise is
that we can then use this extended understanding to demonstrate behavior and adaptability typically seen in humans and
animals. These behaviors include rapid decision capabilities
with autonomously driven goal setting. Such an affect-based
approach also offers promise in providing a new approach to
semantics and other problems, including, eventually, even the
ability to make analogies and be creative [19].
The basis of our earlier work [18] was the neuroscience
model of affect by Dr. Jaak Panksepp. Panksepp [30] postulates the existence of seven core (fundamental) emotional systems that map directly to seven physical neuronal systems that
he had identified in the mammalian brain which generate
well-organized emotion sequences that can be evoked by localized electrical stimulation of the brain. He called these systems the SEEKING, RAGE, FEAR, PANIC, LUST, CARE,
and PLAY systems, with the latter three qualified as serving
socioemotional needs. Panksepp distinguishes his emotions
from basic drives such as for example hunger. Per Pank-
Panksepps more sophisticated, special purpose, socioemotional systems, i.e., PLAY, LUST, and CARE.
Every module of the brain is made up of neurons whose activations are known to usually activate exponentially (see for example Armony et al. 1997, Greenfield, S. 2000). This motivated us to use exponential relationships in this initial attempt
at modeling the different neural systems. These equations are
tentative; place-holders until research from the neurosciences
provide us with the basis for more accurate relationships.
Nonetheless, it appears that even with these heuristic approximations, a useful decision system with seemingly realistic behavior is possible.
As a first approximation, we based the equation we used to
model FEAR, on the distance to the threat. If the Manhattan
distance to the threat is greater than some threshold value, then
the prey takes no evasive action (shows no FEAR). Let this
threshold distance be X F . However once this threshold is
crossed, then the prey exhibits FEAR behavior (avoidance), and
we need a function to model this FEAR. Following from our
earlier discussion on neurons activating exponentially, we
choose an exponential function as an initial, approximate model
for FEARPac-Mans FEAR increases from zero at distance
[greater than or equal to] X F from a ghost, to infinity at a distance of zero from the ghost, as follows:
FEAR = e
XF
)
X
for X
Likewise, we use a similar approach for SEEKING. Starvation studies (Owen and Hanson 2004) indicate that humans
experience hunger pangs a few hours after a meal. These continue for the first three days without food, at which point, the
body switches to start consuming body fat. Once this happens,
the average human feels no further feelings of hunger for approximately five weeks. After this, with body fat consumed,
the system starts consuming muscle. The person then experiences extreme hunger that signals that they must find food or
else, their body will start consuming body muscle and brain
matter to survive. Rather than model this complex behavior in
detail, we again chose as a first approximation an approach
similar to that used with the FEAR dimension, except time
based. If the prey last consumed food at time TS , then the
SEEKING value for Pac-Man goes up exponentially for values of time T greater than TS , until T TS reaches another
value TD , at which time the prey dies from starvation. Like-
SEEKING = e
TD
)
T D ( T T S ) f
*e
XD
)
X
(2)
the prey to boundaries in the x and y directions ( B X , BY respectively), and coming into play only when this value is below
some threshold value X E , i.e.:
XE
(min( B X ) + min( BY ))
(3)
for < X E
sponding
standard
deviation
and
mean
xi
for
The preys affective state is hence contained in the dimensions of FEAR, SEEKING, RAGE and time, i.e., its affective
evaluation of the predator and its food, are represented by the
values of RAGE, FEAR, and SEEKING that result from its interacting with them and its own need for survival and wellbeing. At any given time, these values represent the preys semantic understanding of the predator. The overall affective
state of the prey at any given time is the net of the positive affect of SEEKING and the negative affects of FEAR and
RAGE:
PREYoverall _ affect = SEEKING FEAR RAGE (4)
This research (Joseph and Levkowitz 2012a) made a convincing case that adaptive behavior was possible with just the
simpler three emotions. Once we understand these better, we
plan to incorporate the more complex emotions not presently
considered. Hence, we view this model only as the start of a
program that will grow to include the more sophisticated emotions as knowledge of these become more available.
In this earlier work though, the values of TD , X F , X D , X E in
the equations were hard-coded. Since the results with this
hard coding were promising, we felt justified in going on to the
next stage and evolving these parameter values.
Figure 1: Setup used to evolve agent parameters
i = i (1 + .N (0,1))
xi = xi + i.N (0,1)
(5)
Results were as expected from standard concepts of evolutionary programming: affect parameters changed and evolved
depending on the number of predators and prey, different space
and food configurations of the environment, and with differing
degrees of rates of convergence and stability, and depending on
the population size. In short, as expected, we learned nothing
new about evolution per se. Rather, we found that the agents
were able to use this scheme to adapt to their environment, as
was to be expected. The basic finding was that EP is a useful
technique to use, allowing autonomous agents to learn and
adapt to their environment. We found that average computing
time on a low end 2.3GHz Intel CPU for 1000 runs of the
model was 7.08 seconds, in line with our LeDoux based heuristic of 10 ms., i.e., the model is fast enough for use in real-time
applications.
140
120
FEAR Factor
Affect Value
100
80
SEEK Factor
60
FOOD Factor
40
RAGE Factor
20
0
1
2001
4001
6001
8001
10001
12001
14001
16001
18001
20001
22001
24001
26001
28001
30001
Iteration Number
Figure 3 shows a setup we used to test the ability of our affective systems model to demonstrate this kind of behavior. We
programmed a predator (blue) to march only horizontally,
choosing at random whether to go to the left or to the right, and
thereby creating a guarded boundary. To pass through this
guarded boundary, the prey must expose itself to considerable
danger. We programmed the prey (yellow) with the emotional
system. All the food is on the side of the predator away from
the prey and the prey has to cross the predators line of marching to get to the food.
Initially, we see that the prey stays as far away from the
predator as per its value of FEAR. However, as the time since
it has last eaten increases, its SEEK value goes up and it begins
to move to the food, despite the presence of the predator. As
the predator comes closer to the food, the SEEK value also
goes up and directs it towards the food. At some point the
SEEK overrides the FEAR and the prey moves towards and
crosses the line of action of the predator. The position of the
second piece of food is in the corner: SEEK must exceed
RAGE (a function of closeness to enclosing walls) before the
prey will consume this food.
We found that the values of the parameters as evolved in
the environment of (say) Figure 1 were far from suitable for
such a task. The prey was never able to summon up enough of
SEEK to overcome its FEAR. We had to evolve these values
specifically for the Turing test environment before we saw the
expected outcome of the prey overriding its FEAR, making it to
its food, and thereby surviving. The evolved value of FEAR
for this more aggressive situation where Pac-Man had to pass
the ghost was about 25 times less than in the previous case.
RAGE (a measure of the prey being trapped by the boundaries)
was about five times smaller, signifying that Pac-Man had
evolved more aggressive behavior and so less likely to not approach Pac-Man or to stay away from the boundaries of its environment. Likewise, the evolved values of the prey indicated
NET AFFECT
10000
NET AFFECT (affect units)
5000
0
1
7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73
-5000
-10000
-15000
-20000
-25000
Prey Step Number
of affect, and with rules generated internally based on the interaction of the objects with the general goal of maximizing agent
well-being. As more information of the biological underpinnings of affect becomes available, we surmise that biological
based approaches, presently very much in their infancy, will
eventually match the abilities of current, sophisticated, appraisal theory based approaches.
6. Conclusion
We aim our future work towards generalizing our approach to
handle more complex situations than the Pac-Man game. To do
this, we plan to explore analytical methods such as reinforcement learning to replace the current greedy optimization
strategy with a more efficient approach; introduce the concept
of learning during an agents lifetime based on the localized
damage (health) of a single agent; introduce memory and analytical reasoning; include more emotions such as PANIC so as
to model social situations, such as the teenage boy on his first
date; enrich environment diversity by introducing other affectenabled agents with different characteristics; improve the equations used in the affect model by continuing to study the underlying empirical evidence as it becomes available from mainstream neuroscience.
Rodney Brooks (2002) proposed a subsumption architecture where each layer subsumes the functionality of the layer
below it. In our system, there seems an obvious place for a like
approach: an affect layer subsumes the layer that manages the
agents body, and with a cognitive layer subsuming the affect
layer. This three-layer, triune architecture offers the possibility of a new kind of autonomous agent, one that because of its
ability to develop affect-based understandings of objects based
on its interactions with them is capable of developing semantic
understandings of these objects and hence demonstrates humanlike behavior in social or other complex emotional situations.
7. References
1. Armony, J. L., Servan-Schreiber, D., Cohen, J. D., and
LeDoux, J. E. 1997. Computational modeling of emotion: explorations through anatomy and physiology of fear conditioning. Trends in Cognitive Science, 1, 28-34
2. Barrett, L. 2006. Are emotions natural kinds? Perspectives on Psychological Science 1, 28-58
3. Barrett, L., Mesquita, B., Ochsner, K. N., and Gross, J. J.
2007. The experience of emotion. Annual Review of Psychology, 58, 373-403
4. Blumberg, B., Todd, P., and Maes, P. 1996. No bad dogs:
ethological lessons for learning. In: Proceedings of SAB96.
Cambridge, MA: The MIT Press
5. Bratman, M. E. 1987. Intention, Plans and Practical Reason Stanford, CA: CSLI Publications
6. Canamero, D. 1997. Modeling motivations and emotions
as a basis for intelligent behavior. In W. Lewis Johnson, ed.,
Proceedings fo the First International Conference on Autonomous Agents, New York, NY: ACM Press, 148-155
7. Canamero, D. 1998. Issues in the design of emotional
agents. In emotional and intelligent: the tangled knot of cognition. Papers from the 1998 AAAI Fall Symposium, Technical
Report FS-98-03, Meno Park, CA: AAAI Press, 49-54
8. Combes, S. A., Rundle, D. E., Iwasaki, J. M., and Crall, J.
D. 2012. Linking biomechanics and ecology through predator